Research map / Adversarial robustness and security / Black-box and transfer attacks
Cross-Domain Transferability of Adversarial Perturbations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use relativistic cross-entropy instead of standard cross-entropy (ablation) | Adversarial perturbation generator trained on Paintings · ImageNet val-set, transfer from ResNet-152 to VGG-16 | fooling rate (%): 79.21 -> 86.95 | Helped |
| Use relativistic cross-entropy instead of standard cross-entropy (ablation) | Adversarial perturbation generator trained on Paintings · ImageNet val-set, transfer to VGG-19, Squeeze-v1.1, and Dense-121 | fooling rate (%): 78.96 -> 85.88; 69.32 -> 77.81; 66.45 -> 75.21 | Helped |
| Train the generator on Comics rather than using Gaussian noise | Cross-domain black-box attack trained against ChexNet · ImageNet val-set, ε≤10, VGG-19 | fooling rate (%): 23.59 -> 48.47 | Helped |
| Train the generator on Comics rather than using Gaussian noise | Cross-domain black-box attack trained against ChexNet · ImageNet val-set, ε≤10, Dense-121 | fooling rate (%): 17.05 -> 31.81 | Helped |
| Add Gaussian smoothing to generator output (ablation) | Generator trained on Paintings · ImageNet-NeurIPS, adversarially trained Inc-v3 ens3 | success rate (%): 69.3 -> 85.2 | Helped |
| Add Gaussian smoothing to generator output (ablation) | Generator trained on Paintings · ImageNet-NeurIPS, naturally trained IncRes-v2 | success rate (%): 99.8 -> 97.6 | Hurt |
| Train for more epochs (ablation) | Generator trained on Paintings · IncRes-v2 transferability | transferability | Mixed |
| Increase Gaussian kernel size (ablation) | Smoothed generator output · IncRes-v2, naturally and adversarially trained | transferability | Hurt |
| Train on Paintings, Comics, or ChestX instead of target-domain ImageNet data | Universal adversarial generator · ImageNet val-set cross-domain black-box attack | fooling rate: up to 88.12 | Helped |
Closest papers by meaning
- Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks (CVPR 2019 · similarity 0.92)
- Learning Transferable Adversarial Perturbations (NeurIPS 2021 · similarity 0.91)
- Boosting the Transferability of Adversarial Attacks with Reverse Adversarial Perturbation (NeurIPS 2022 · similarity 0.91)
- Towards Transferable Adversarial Attacks with Centralized Perturbation (AAAI 2024 · similarity 0.91)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.90)
- Boosting Adversarial Transferability through Augmentation in Hypothesis Space (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.